Current work

Expertise and Trust in Humanities AI

What AI changes about humanities research, and who has to be trusted for any of it to count.

Fred Gibbs: fredgibbs.net
History: history.unm.edu
Amaranth: amaranth.unm.edu

Amaranth

A studio, not a service desk.

Amaranth is UNM’s digital humanities studio: a room, a printer, a few machines, and students. No dev team, no grant-funded engineering staff. Faculty are asked to bring a half-formed question rather than a finished project, and the work gets built in public on infrastructure the authors keep.

amaranth.unm.edu

Built by undergraduates and one faculty member.

Research

We can already trust AI.

Do you trust me? That is the real issue. AI is trustworthy in the sense that skilled users know what to expect and that AI output is already consistent (while improving). The question we need to answer is how do we know to trust those who use it?

1:36–2:18

Now the problem, and it is not the one the room expects. This is also not a position I arrived with — it is what the studio work kept running into, project after project, until it became the research question.

I am not much worried about whether the model is reliable. Inside a task you have scoped, it is: a skilled user knows what it will do, the output is consistent, and it improves. Treating “can we trust AI” as the open question has eaten a lot of faculty meetings and settled nothing.

The open question is the one on the slide. When I hand you a transcript, a map, a reconstruction, a student essay, you are not evaluating the model. You are evaluating me — whether I knew enough to check it, whether I actually checked, and whether I told you where I stopped. That is a claim about expertise, and we have no established way to warrant it yet.

So: do you trust me? Everything after this slide is the attempt to earn a yes.

AI Sketchbook

Field notes, including the failures.

A shared space where colleagues try things and compare notes — fifteen sketches so far, each tagged with whether it was actually tested.

2:18–2:48

This is where the research is actually happening, and it is deliberately unglamorous. Colleagues try something, write down what happened, and tag it. Each entry has a status — tested, untested — because the honest answer is usually “this half worked.”

The reason it is a website and not a paper: the useful unit here is a small, dated, local observation, and by the time it is a paper it is out of date and no longer local.

Read the last line of the intro out loud: students should stay alert to “the borrowed, uneven, and sometimes misleading expertise AI seems to offer.” Borrowed expertise is the whole problem from the previous slide in four words — and noticing when you are holding some is a skill, which means it can be taught and it can be evidenced.

Research workflows

Scale is no longer the obstacle.

Reading across thousands of pages, drafting transcripts for a whole interview collection, describing a few hundred photographs so they can be searched — all of it is now a few weeks of supervised work rather than a funded project. Which means the binding constraint is no longer capacity. It is judgment, and we have not built the training for that.

2:48–3:30

The first focus: bringing AI into humanities research itself, not just into the classroom.

Be concrete about what changed. It isn’t that the machine is smart. It’s that three or four specific bottlenecks — transcription, description, normalization of messy archival data, first-pass reading at scale — got cheap enough that a single researcher or a class can now attempt a collection that used to need a grant and a team.

And be equally concrete about what didn’t change: every one of those outputs is wrong somewhere, usually in the places that carry the most meaning. So the work didn’t disappear; it moved to checking, and checking requires more expertise than producing did.

Next on this list is mapping: georectification, layer alignment and feature extraction are exactly the technical drudgery that keeps spatial analysis out of a humanities course. AI does the lift; the reading stays ours.

Design sketch · nothing is built

MemoryTour

Oral history, indexed by place.

A design sketch drawn for this talk. The transcript is machine-made and marked unchecked, because that is the state it would actually be in.

3:30–4:15

This is the shape of the thing rather than the thing — say that out loud, and say it before anyone has to ask.

What it is for: we have interview collections whose content is overwhelmingly about places — a ditch, a field, a building, a corner — and we store them as a list of files named after people. Indexing by place instead means you can stand somewhere and hear who has talked about it.

The reason this is newly possible: draft transcription across a whole collection is now cheap, and you can get first-pass place extraction out of the same pass. The reason it is still hard is on screen: that transcript is unchecked, and machines mangle exactly the words a project like this is about — Spanish, family names, irrigation vocabulary, local place names. So the label stays on until a person has listened.

If the demo is working, switch to it here.

Reconstruction

Every reconstruction is an argument with the confidence turned up.

LiDAR and photogrammetry fail visibly: a gap in a point cloud looks like a gap. A generated elevation looks finished whether or not anyone alive knows what was there. If we build these, the drawing convention has to carry the evidence — which is a discipline architectural illustration has had for a century and the new tools do not.

Simulation · not a working app Arcade · photographed Windows · inferred Roofline · no source

Alvarado Hotel · 1902–1970

Solid is photographed. Dotted is invented.

Simulated AR view, drawn for this talk. The line weight is the evidence, the way it is on an archaeological site plan.

4:55–5:45

Spend your time here. This is a drawing of a proposal, not a product — I would rather show an honest drawing than a dishonest render.

The building is the Alvarado Hotel: Charles Whittlesey for the Santa Fe Railway, Mission Revival, the largest of the Harvey hotels, demolished in 1970 over citizen protest. Everyone in Albuquerque knows the hole where it was and almost nobody under sixty has stood inside it.

Now point at the line weights. Solid gold is the arcade, which is extensively photographed. Dashed is the second storey, inferred from what buildings of this type did. Dotted is the roofline and the towers, where I would be inventing. The interface shows its own epistemics — and the confidence lives in the drawing convention, not in a disclaimer nobody reads.

Sources: Charles Whittlesey, Santa Fe Railway, Mission Revival, demolition 1970 — Wikipedia, “Alvarado Hotel.” Confirm before presenting if any detail is load-bearing.

Teaching

The site is the deliverable. The paper never was.

When public scholarship is the default output rather than something translated out of a seminar paper afterwards, the incentives change on day one: students write for readers, cite for strangers, and design for someone arriving by search. Research nobody can find is research that didn’t happen.

UNM Campus Histories

A campus surveyed by the people who live in it.

A multi-semester project. Each cohort builds on what the last one made, and the categories on the essay index were invented by students who needed to organize their own work.

Community storytelling

The knowledge is already in the community.

Farmers along the Middle Rio Grande hold knowledge about water, soil, and land use that exists in no document anywhere. Two courses — qualitative methods and local food systems — went and asked. The site is where what they said becomes citable.

What actually changed

Draft transcripts in days, not months.

Transcription used to be the budget line that killed the project. It is now the cheap part — and the part that needs checking hardest, because machines mangle names, Spanish, and every term specific to a place.

Stewardship

Not everything belongs in the nearest tool.

Community oral histories, Indigenous collections, unpublished archives: the material that most rewards this work is exactly the material you cannot paste into whatever service happens to be open. Before anything is uploaded we ask who has a stake in it, what the permissions actually cover, and which tool fits — including keeping it on a machine in the room. That conversation is the method, not the paperwork.

Farmer profiles

Narrator, interviewer, author — all named.

Every profile credits the farmer, the student who ran the interview, and the students who wrote it up. Attribution is the visible half of stewardship; the rest happens before the recorder goes on.

Infrastructure

Every part of a page has a name.

All of this runs on Xanthan: plain text files, Jekyll, GitHub Pages — free to host, nothing to renew, and no vendor who can raise the rent or switch it off. Naming the parts is what lets an AI change one thing instead of rewriting the site, and what lets a student see exactly what changed. A project that outlives its funding is a preservation decision taken on day one.

9:22–10:05

This is the piece that makes the rest of the talk reproducible rather than a story about one lucky studio, so give it a moment.

Two claims. First, durability: plain text on open standards means a project outlives its grant cycle and its author’s interest. Some of the sites in the gallery you’re about to see have not been touched in years and still work.

Second, and this is the one I care about: the framework is built out of named parts. That constraint is what makes AI assistance safe to hand to a nineteen-year-old. You ask for a change to one named block, you get a change to one named block, and you can see it in the diff. The failure mode of AI web work — ask for a tweak, get a plausible rewrite of everything — is designed out at the level of the building blocks rather than patched at the level of the prompt.

xanthan-web.github.io

All of it is plain text underneath.

Course archives, oral history collections, digital exhibits, scrolling narratives, mapped directories, annual reports — every one of these is a real site made by UNM students and faculty.

3D pedagogy

Scan it — then put it back in a hand.

We printed five medieval capitals from Scan the World’s collection. Set them on a table and the Romanesque-to-Gothic transition stops being a claim in a textbook: it becomes something you see by turning two objects. Scanning is the middle of the pipeline, not the end of it.

10:35–11:12

A short detour into 3D, because it is the same argument in a different material.

The files are free and public — Scan the World hosts scans from museum collections. Printing five of them costs almost nothing. What you get for that is a comparison a student can make with their hands, and the print quality is limited by the scan in ways that are themselves worth discussing in class.

The point for this talk: a digitized object that stays on a screen is a photograph with extra steps. The pedagogical value shows up when it comes back out into the room.

The partnership version of this: students have been scanning artifacts at UNM’s Maxwell Museum and documenting the workflow so the next person can repeat it, and student researchers have taken 3D printing into high school history classrooms. An undocumented digitization is a one-off, not infrastructure.

Amaranth · 3D pedagogy

The objects on a table, with the argument printed beside them.

Capitals poster, Amaranth studio. Five prints from Scan the World, Romanesque to Gothic, each captioned with its date and the museum that holds the original.

Speculative fiction

Ask it for a place that never existed.

The most useful exercise I have found puts all of this together: have a model generate a landscape, a building, a settlement — something plausible and wholly invented — and then argue about it. The image is not the assignment. The argument is the assignment.

AI-generated image

What here is real? What isn’t?

A generated model landscape of cliff dwellings. No site, no survey, no scale. Students are asked which parts they would defend, on what evidence, and what would settle it.

12:25–13:30

Put it up and let the room work on it for a moment before you say anything. Say clearly that it is AI-generated — after the last two slides it would be absurd not to.

Then the questions, in order. What is doing the work of realism here? The contour terrain reads as a survey model, which is a claim about measurement that nothing in this image is entitled to make. Where are the dwellings placed, and is that where they are actually built? What is the construction — is that masonry coursing or is it texture? What is missing that would be there: middens, water, paths, agriculture, any sign of how anyone got in or out?

A student who knows the Southwest will get to “the siting is wrong” fast. Getting them to say how they know is the entire pedagogical payload.

The point

The complaint is the lesson.

“None of this is real” is exactly the response the exercise is built to produce — and usually the first time a student has had to say precisely why it isn’t, and what evidence would settle it. That is the skill I am trying to teach, and it transfers to every other confident thing a machine hands them.

Fred Gibbs · fredgibbs.net · amaranth.unm.edu

Department of History & Amaranth, University of New Mexico